About

Suleman Surti, PhD, is a Research Professor of Radiology at the University of Pennsylvania's Perelman School of Medicine. He is a full member in the Radiobiology and Imaging Program at the Abramson Cancer Center. His research expertise includes PET imaging, instrumentation, system modeling, image generation, and image analysis. Dr. Surti's work focuses on advancing imaging techniques and systems within the field of radiology, contributing to the understanding and development of medical imaging technologies.

Research topics

• Physics
• Optics
• Computer science
• Nuclear medicine
• Artificial intelligence

Selected publications

• A Virtual Clinical Trial to Detect Changes in Tumor Uptake with PET using Lesion Embedding

  IEEE Transactions on Radiation and Plasma Medical Sciences · 2026-01-01

  articleOpen accessSenior author

  ), and are relatively independent of the local background (organ) uptake. The primary factor determining the AUC values seems to be the absolute change in lesion uptake that will be sensitive to partial volume effects as determined by the scanner spatial resolution. Hence, a standard axial field-of-view scanner with similar spatial resolution will likely perform as well as a LAFOV scanner after appropriate compensation for the sensitivity differences.

  PublisherOA PDFDOI

• Sensitivity-Weighting Methods in Reconstruction and Denoising for Axially Sparse Long Axial Field-of-View PET: A Lesion Embedding Study

  2025-11-01

  article

  The emergence of long axial field-of-view (AFOV) PET scanners offers unprecedented sensitivity, enabling better image quality, lower dose, and faster scans. As evidence of their clinical value grows, adoption of these scanners is expected to rise, but the optimal AFOV still depends on the intended clinical application, balancing imaging performance against system cost. Deep learning denoising (DL-DN) techniques may further support cost-effective designs to optimize performance even with shorter or sparsely configured AFOV systems. However, system designs with nonuniform sensitivity profiles lead to noise variations and reduced lesion detectability in lowsensitivity regions. While current DL-DN models are effective at reducing image noise, they are not trained with explicit knowledge of sensitivity variations. Focusing on a sparse long AFOV PET design with axial gaps, we investigate the potential benefit of a sensitivity-informed DL-DN network to tailor image quality enhancement for low- and high-sensitivity regions. We used a sensitivity-weighted (SW) OSEM image reconstruction and a phantom study on the PennPET Explorer to test how incorporating sensitivity knowledge could improve noise uniformity and achieve consistent lesion detectability performance across the AFOV, comparing a fully-populated long AFOV design and an axially sparse system configuration. The current (non-sensitivityinformed) DL-DN model fails to preserve lesion contrast and detectability in low-sensitivity (gap) regions, whereas SW-OSEM reconstruction provides more stable, noise-uniform reconstructions.

  PublisherDOI

• Depth-of-interaction encoding techniques for pixelated PET detectors enabled by machine learning methods and fast waveform digitization

  Physics in Medicine and Biology · 2025-04-04 · 1 citations

  articleOpen access

  Abstract Objective . Pixelated detectors with single-ended readout are routinely used by commercial positron emission tomography scanners owing to their good energy and timing resolution and optimized manufacturing, but they typically do not provide depth-of-interaction (DOI) information, which can help improve the performance of systems with higher resolution and smaller ring diameter. This work aims to develop a technique for multi-level DOI classification that does not require modifications to the detector designs. Approach . We leveraged high-speed (5 Gs s −1 ) waveform sampling electronics with the Domino Ring Sampler (DRS4) and machine learning (ML) methods to extract DOI information from the entire scintillation waveforms of pixelated crystals. We evaluated different grouping schemes for multi-level DOI classification by analyzing the DOI positioning profile and DOI positioning error. We examined trade-offs among crystal configurations, detector timing performance, and DOI classification accuracy. We also investigated the impact of different ML algorithms and input features—extracted from scintillation waveforms—on model accuracy. Main results . The DOI positioning profile and positioning error suggest that 2- or 3-level binning was effective for 20 mm long crystals. 2-level discrete DOI models achieved 95% class-wise accuracy and 83% overall accuracy in positioning events into the correct DOI level and 3-level up to 90% class-wise accuracy for long and narrow crystals (2 × 2 × 20 mm 3 ). Long short-term memory networks trained with time–frequency moments were twice as efficient in training time while maintaining equal or better accuracy compared to those trained with waveforms. Classical ML algorithms exhibit comparable accuracy while consuming one order less training time than deep learning models. Significance . This work demonstrates a proof-of-concept approach for obtaining DOI information from commercially available pixelated detectors without altering the detector design thereby avoiding potential degradation in detector timing performance. It provides an alternative solution for multi-level DOI classification, potentially inspiring future scanner designs.

  PublisherDOI

• Self-Supervised Attention-Based Deep Learning for Cardiac Motion Correction in Dynamic PET Imaging

  2025-11-01

  articleSenior author

  Motion artifacts remain a challenge to accurate quantification in dynamic cardiac PET imaging, where traditional gating and motion correction methods often suffer from inefficiencies, low robustness, and sensitivity to irregular motion. This study introduces a novel self-supervised, attention-based deep learning framework that directly estimates motion between PET time frames without requiring external gating signals or motion vector fields. The model integrates intra- and inter-frame attention mechanisms to learn both local and global motion patterns, enabling accurate spatial alignment while preserving regional tracer uptake–critical for kinetic modeling. Using synthetically generated dynamic PET data with realistic cardiac motion and including Poisson noise to mimic low-count frames, the model achieved near-perfect structural alignment (SSIM <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\approx 1.0$</tex>) and maintained uptake fidelity, with uptake errors under 1% across all time frames. These results demonstrate the potential of a gating-free, attention-based approach for motion correction in dynamic PET, with future work focused on adapting to clinical data and extending to whole-body applications.

  PublisherDOI

• Quantitative Evaluation of Spatially-Variant Deformations Recovered by Deep Learning on Clinical-Like Breast Lesions

  IEEE Transactions on Radiation and Plasma Medical Sciences · 2025-01-01

  articleOpen access

  The dedicated dual-panel breast PET scanner (B-PET) can potentially provide improved imaging of breast lesions with higher spatial resolution and sensitivity and lower scanner costs. On the other hand, the dual-panel systems suffer from spatially-variant deformations, due to the limited angular sampling and parallax errors, that need to be mitigated in order to improve visualization and quantification of the breast lesions. Our previous studies demonstrated that the deep-learning approach can significantly reduce spatially-variant deformations in the context of B-PET. However, these studies used only simplified geometric objects with a uniform background for training and testing, which does not fully represent the clinical complexity in breast PET imaging. In this work, we developed a methodology for generation of synthetic clinical-like breast images with complex lesion shapes, including spiculated lesions and following tracer-dependent activity characteristics for 18FFES and 18FFDG. Subsequently, the neural network was trained and tested on clinical-like synthetic B-PET data reconstructed using statistical iterative reconstruction (DIRECT-RAMLA) used as the network input. Our results show that the deep learning approach can substantially suppress deformations for both small lesions and the larger phantom itself in dual-panel PET reconstructions. The deep learning approach can also improve quantitative measurements in terms of the lesion contrast metrics and image roughness over the phantom background. A research 18FFES patient study further confirms the improved visual image quality and a good visual correlation for a heterogeneous lesion with the total body PET imaging result.

  PublisherOA PDFDOI

• A Task-Based Evaluation of Lesion Detectability in a Deep Learning Reconstruction Algorithm for PET

  2025-11-01

  articleSenior author

  FastPET<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[1]-[3]</sup> is a novel neural network-based reconstruction that is computationally efficient for fully-3D PET reconstruction using a simple architecture with the fewest network parameters. Clinical task-based evaluations are critical to enable the application of FastPET in clinical practice. In this work, we study lesion quantification and detectability, key clinical tasks in oncology, for FastPET reconstruction relative to the conventional OSEM reconstruction. Patient data acquired from Biograph Vision and a well-developed lesion embbeding methodology<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[4],[5]</sup> were used. Lesions were embedded at different uptake levels into patient data for varying scan durations. Results show comparable lesion contrast measurements for FastPET and OSEM images. For a fixed scan time, FastPET also has a narrower background nodule contrast distribution that is consistent with a visually smoother background. Detailed evaluation of lesion contrast recovery and lesion detection and localization as characterized by the area under the localization receiver operating characteristic curve (ALROC) metric will be presented as a function of scan time and lesion uptake value. Such task-based evaluation of small-lesion detection and quantification will provide clinical significance and evidence of introducing Deep Learning reconstruction for PET in clinical practice.

  PublisherDOI

• Total-Body PET System Designs with Axial and Transverse Gaps: A Study of Lesion Quantification and Detectability

  Journal of Nuclear Medicine · 2025-01-16 · 1 citations

  articleOpen accessSenior author

  High-sensitivity total-body PET enables faster scans, lower doses, and dynamic multiorgan imaging. However, the higher system cost of a scanner with a long axial field of view (AFOV) hinders its wider application. This paper investigates the impact on the lesion quantification and detectability of cost-effective total-body PET sparse designs. <b>Methods:</b> Using the PennPET Explorer (PPEx) as a model, 3 sparse configurations with the same 142-cm AFOV were considered, including designs with only axial gaps (AGs), only transverse gaps (TGs), and a mixture of AGs and TGs (MG), with retained detector fractions (DFs) ranging from 71% to 40%. Human data from the PPEx were used to emulate sparse designs by discarding lines of response as a proxy for missing detectors. We embedded lesion events in the resultant list data with varying uptakes in the lung and liver before reconstruction. A generalized scan statistics methodology was used to measure lesion detectability and quantification as a function of lesion uptake and scan duration. <b>Results:</b> Relative to a fully populated system, an AG design with 71% performs well but is susceptible to image artifacts as the DF decreases to 58%. A TG design performs well with a DF of 58% but requires twice the scan time to achieve similar lesion detectability and is susceptible to transverse field-of-view truncation below 60 cm as the DF is further decreased. An MG design with a DF of 58% requires 3 times the scan time to achieve similar lesion detectability, and with no evidence of artifacts even as the DF is decreased to 40%. <b>Conclusion:</b> Sparse designs with artifact-free images can provide comparable lesion quantification and detectability to the fully populated PPEx after compensating for the reduced sensitivity with increased scan time. Because an AG design is more susceptible to image artifacts with a lower DF, a system with only AGs is not an optimal choice for dramatic cost reduction. A TG design provides a higher relative sensitivity than AG or MG designs for a given DF, leading to a shorter scan time to achieve comparable lesion detectability. However, the increased truncation of the transverse field of view with decreasing DF limits this design choice. An MG design allows for the greatest cost reduction (lowest DF) if the scan duration is increased to compensate for the higher loss in sensitivity. Sparse designs of PET with a long AFOV provide a technologic solution for introducing such systems at reduced cost into routine clinical use.

  PublisherOA PDFDOI

• Impact of Detector Parameters and Image Resolution Modeling on Dedicated Brain PET Imaging

  IEEE Transactions on Radiation and Plasma Medical Sciences · 2025-01-01

  articleOpen accessSenior author

  High performance brain PET scanners suitable for accurate and precise measurement of uptake in small regions of the brain require high spatial resolution (1.5-2 mm) together with high system sensitivity. While small cross section crystals are needed to achieve high spatial resolution with pixelated detectors, effective system sensitivity requires long crystals and good time-of-flight (TOF) resolution. In addition, depending on crystal thickness, some level of depth-of-interaction (DOI) capability may also be needed to reduce parallax error. While various methods for DOI measurement can be utilized within the detector, the impact on reconstructed spatial resolution as a function of crystal thickness needs to be carefully evaluated. The focus of this work is to use Monte Carlo simulations to evaluate the impact of crystal size, thickness, and DOI resolution on the reconstructed spatial resolution in a high-resolution brain scanner design. Two crystal cross sections were chosen, 1.6 mm × 1.6 mm and 2 mm × 2 mm, and three different LSO crystal thickness (10 mm, 15 mm, and 20 mm) were evaluated. We evaluated these system designs without DOI capability and with DOI resolution (4-12 mm) for varying crystal thickness. In addition, we evaluated the use of image-space resolution modeling (IRM) techniques, which can mitigate losses in spatial resolution and positioning accuracy due to parallax error and Compton scatter, and to complement the advantages of DOI capability. A brain size elliptical phantom with small spheres was simulated to study the overall benefit in uptake measurements for small objects using the contrast recovery coefficient (CRC) metric. Our results show that a system with 2 mm × 2 mm crystals can provide comparable image quality in terms of CRC measurement than one with 1.6 mm × 1.6 mm crystals. However, crystal thickness has a bigger impact on spatial resolution and CRC with performance degrading with best performance achieved with 10 mm thick crystals. DOI measurement is necessary with longer crystals in order to maintain good imaging performance with high sensitivity. For the 15 mm thick crystals 9 mm DOI resolution is needed to compensate for the loss in spatial resolution compared to 10 mm thick crystals without DOI, but still with lower CRC. A similar conclusion is reached for 20 mm crystals with 4 mm DOI resolution. Alternatively, IRM methods can provide improved and similar performance for all crystal thicknesses studied here, indicating a pathway for achieving our brain scanner design goals without additional cost, complexity, and performance tradeoffs. This work provides a systematic investigation for designing future brain PET systems.

  PublisherOA PDFDOI

• Recovery of the spatially-variant deformations in dual-panel PET reconstructions using deep-learning

  Physics in Medicine and Biology · 2024-02-08 · 6 citations

  articleOpen access

  Dual panel PET systems, such as Breast-PET (B-PET) scanner, exhibit strong asymmetric and anisotropic spatially-variant deformations in the reconstructed images due to the limited-angle data and strong depth of interaction effects for the oblique LORs inherent in such systems. In our previous work, we studied time-of-flight (TOF) effects and image-based spatially-variant PSF resolution models within dual-panel PET reconstruction to reduce these deformations. The application of PSF based models led to better and more uniform quantification of small lesions across the field of view (FOV). However, the ability of such a model to correct for PSF deformation is limited to small objects. On the other hand, large object deformations caused by the limited-angle reconstruction cannot be corrected with the PSF modeling alone. In this work, we investigate the ability of deep-learning (DL) networks to recover such strong spatially-variant image deformations using first simulated PSF deformations in image space of a generic dual panel PET system and then using simulated and acquired phantom reconstructions from dual panel B-PET system developed in our lab at University of Pennsylvania. For the studies using real B-PET data, the network was trained on the simulated synthetic data sets providing ground truth for objects resembling experimentally acquired phantoms on which the network deformation corrections were then tested. The synthetic and acquired limited-angle B-PET data were reconstructed using DIRECT-RAMLA reconstructions, which were then used as the network inputs. Our results demonstrate that DL approaches can significantly eliminate deformations of limited angle systems and improve their quantitative performance.

  PublisherOA PDFDOI

• Correlation of NEC to Image SNR and Lesion Detectability in Clinical PET

  2024-09-25

  articleSenior author

  In positron emission tomography (PET) imaging, Noise Equivalent Counts (NEC) serve as a key indicator of overall system performance. Previous studies have established the relationship between NEC and signal-to-noise (SNR) for both non-time-of-flight (non-TOF) and TOF PET scanners employing filtered back-projection (FBP) algorithms. This study aims to explore the correlation between NEC and image SNR, as well as task-specific metrics, using iterative reconstruction with comprehensive TOF PET system modeling. The phantom, 35 cm in diameter to represent a patient with $\mathrm{BMI}\gt30$, featured 12 1-cm diameter spheres placed in an off-centered slice, and data were acquired across a wide range of activity levels using the PennPET Explorer. Multiple scans were generated via bootstrapping and reconstructed using list-mode TOF-ordered subset expectation maximization (TOF-OSEM) algorithms to obtain mean and standard deviation images. Image SNR was quantified as the ratio of mean values within a central volume-of-interest (VOI) across both images. Utilizing a generalized scan-statistic model, the area under the localization receiver operating characteristic (ALROC) was calculated for the spheres placed at the offset slice. NEC was computed using both conventional and TOF-adapted formulas from the original data. The results indicate that despite the linearity between SNR and NEC metrics, neither conventional NEC nor $\mathrm{NEC}_{\text {TOF }}$ metrics accurately represent the global image SNR when employing iterative reconstruction methods. Furthermore, our results show that although SNR is proportional to ALROC, it does not adequately represent the change in ALROC as TOF resolution improves.

  PublisherDOI

Recent grants

• NIH Grant R01EB009056

  NIH · $1.5M · 2015

• Optimization of Clinical and Research PET Imaging

  NIH · $2.2M · 2020–2025

• High Performance, Quantitative Breast PET Scanner Integrated With Tomosynthesis

  NIH · $5.9M · 2016–2027

• Time-of-Flight PET for Improved Whole-Body Imaging

  NIH · $9.8M · 2006–2026

• NIH Grant R21EB008142

  NIH · $479k · 2011

Frequent coauthors

• Joel S. Karp

  Institut National des Sciences Appliquées de Toulouse

  148 shared

• Margaret E. Daube-Witherspoon

  University of Pennsylvania

  27 shared

• G. Muehllehner

  21 shared

• Amy E. Perkins

  Philips (United States)

  18 shared

• Srilalan Krishnamoorthy

  University of Pennsylvania

  17 shared

• Samuel Matej

  University of Pennsylvania

  17 shared

• Matthew E. Werner

  University of Pennsylvania

  15 shared

• Jeffrey P. Schmall

  10 shared